Three STT Strategies, One Market

The Gap Between STT Models Had Nearly Closed

By June 2026, picking a speech-to-text provider based on accuracy alone was a fool’s errand. The STT market competition had squeezed the gap between top providers to the point where the remaining error rate was split between genuinely hard audio (cross-talk, heavy background noise, strong accents) and the diminishing returns of benchmark optimisations that did not transfer to production. Deepgram Nova-3, AssemblyAI Universal-2, Google Chirp, and OpenAI Whisper all landed in the same quality band for clean English audio. The word-error-rate shootouts were down to fractions of a percent that mattered in a pitch deck but not in a real conversation.

This was not true two years earlier. In 2024 there was a clear accuracy ladder: Deepgram and AssemblyAI were ahead of the hyperscalers for conversational speech, and Whisper was a decent self-hosted fallback if you could manage the infrastructure. By 2026 the performance gap had compressed. Every modern STT model transcribed a clear phone call well enough. The question was no longer which provider was most accurate. The question was what else the provider could do with that accuracy.

What Changed: Deepgram Flux Multilingual

The most consequential STT release of 2026 landed on April 29 with Deepgram Flux Multilingual. It was the first single-model conversational STT that supported ten languages simultaneously with native code-switching. That last detail matters more than the language count. Most multilingual STT before Flux required you to pick a language per request or per session, which broke as soon as a bilingual speaker code-switched mid-sentence. Flux handled it automatically, detecting the language shift and transcribing in the correct language without a round-trip through a language identification model.

The supported set was English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch. Not the widest coverage on paper. Google’s Chirp claimed over 100 languages. But Google’s offering used a language-per-channel model under the hood, which meant switching languages required reconfiguring the stream. Deepgram’s single-model approach was architecturally simpler for developers building multilingual voice agents. You opened one WebSocket and sent audio. The model figured out the rest.

Flux also embedded turn detection directly into the recognition model. Previous STT systems relied on external voice activity detection and silence-timeout heuristics to guess when a caller had finished speaking. Flux exposed structured turn-state events: StartOfTurn, EndOfTurn, TurnResumed. For voice agent builders this was a big deal because it removed a whole class of brittle configuration from the pipeline. You set an end-of-turn threshold and the model told you when the other person had stopped talking.

How the Competitive Field Responded

AssemblyAI took a different approach. Universal-2 was a solid general-purpose STT model, but AssemblyAI’s real bet was on their Voice Agent API launched in April 2026. Priced at $4.50 per hour all-inclusive, it bundled STT, LLM orchestration, TTS, and turn detection into a single WebSocket. The pitch was simplicity: one integration, one price, one provider for the whole pipeline. You did not need to assemble components from different vendors. AssemblyAI would handle the lot.

Google Cloud Speech-to-Text kept the traditional hyperscaler strategy: breadth over depth. Chirp covered 100 plus languages, the V2 API had the richest set of integrations in the market (BigQuery, Contact Center AI, Vertex AI), and the pricing scaled aggressively at high volumes. Google’s play was not about having the best single model for any particular use case. It was about being the safest default for any organisation already running on GCP. If you were already spending six figures on Google Cloud, adding STT was a checkbox feature, not a procurement decision.

OpenAI Whisper remained a factor at the self-hosted end of the market. It was free, it ran on your own hardware, and its accuracy was competitive for clean audio. But Whisper had no official streaming endpoint, no turn detection, no built-in speaker diarisation in the open-weight version. Teams that picked Whisper were betting on their ability to build the surrounding pipeline themselves. That was viable for well-resourced engineering teams and a non-starter for everyone else.

Pricing Told the Same Story

The pricing surface made the strategic differences obvious. Deepgram charged by the minute with add-on tiers for diarisation, redaction, and keyterm prompting. Flux English was $0.0065 per minute streaming, Flux Multilingual $0.0078. The pricing rewarded teams that used the raw STT layer and composed their own stack above it.

AssemblyAI charged by the audio hour and itemised every intelligence feature as its own add-on. Voice Agent API pricing was a deliberate departure: one bundled rate for the whole pipeline. This suited teams that wanted a single vendor and were willing to trade granular cost control for operational simplicity.

Google priced by per-minute tiers that dropped sharply at high volume, with separate rates for different model classes (standard, medical, telephony, etc). The architecture customer who committed to Google would outspend everyone on raw STT but get the best unit economics at scale.

Flux Multilingual sat slightly above Nova-3 streaming in price. The premium reflected the model-native turn detection and automatic language switching. AssemblyAI’s bundled Voice Agent API was more expensive than any individual STT line item but cheaper than stitching three separate vendors together. Google’s volume pricing was uncompetitive at low usage and aggressive above a million minutes a month.

The Real Differentiator Was Developer Experience

The accuracy battle was over. The pricing battle was complicated because different buyers optimised for different totals. The one real differentiator that was not a commodity by June 2026 was developer experience and platform integration.

Deepgram had the strongest SDK story among dedicated STT providers. The official SDKs covered TypeScript, Python, Go, Rust, .NET, and Java, with migration guides for teams coming from AWS Transcribe, Google STT, AssemblyAI, and Whisper. Twilio had just integrated Flux Multilingual into Conversation Relay, which meant any developer on Twilio could add multilingual STT with a configuration change rather than a new integration.

AssemblyAI leaned into documentation quality and reference architectures. Their comparison pages against competitors were detailed, their streaming guide was thorough, and the Voice Agent API came with sample apps in multiple frameworks. They were competing on the quality of the hand-holding, not just the quality of the models.

Google had the integration moat. STT plugged into Contact Center AI, Vertex AI agent builder, and BigQuery for post-call analytics. For a developer already on Google Cloud, adding STT was a one-hour task. For a developer not on Google Cloud, the onboarding friction was higher because the docs assumed GCP context.

Who Won Depended on Who You Were

There was no universal winner in the June 2026 STT market. Deepgram was the best choice if you were building a voice agent that needed turn detection and multilingual support from a single model. AssemblyAI was the best choice if you wanted an all-in-one voice pipeline without stitching components together. Google was the best choice if your organisation already ran on GCP and you wanted the integration surface.

This fragmentation was not a bug. It was the market maturing. A commoditised accuracy layer meant the competition moved up the stack to developer experience, platform integration, and the specific shape of the API. The providers that understood which developer persona they were optimising for would keep growing. The ones that tried to be everything to everyone would watch their niche get picked apart.

What This Meant for Voice Agents

For teams building voice agents in production, the narrowing accuracy gap was unambiguously good news. It meant the STT decision was no longer a bet-the-project choice. You could start with one provider and switch to another if the use case demanded it. The migration guides existed, the integration points were standardised around WebSocket streams, and the accuracy floor was high enough that you would not discover a catastrophic failure after shipping.

The real work shifted to the orchestration layer. Turn detection, interruption handling, multilingual routing, and compliance boundaries were now the hard problems. Deepgram Flux solved the first two in the model itself. AssemblyAI solved all of them in the Voice Agent API. Google solved them through integrations with the rest of GCP. The choice was which tradeoffs matched your team’s priorities.

I spent most of my time at Deepgram working with teams that were building voice agents. The teams that shipped fastest were the ones that stopped agonising over model selection and started building the pipeline around it. By June 2026, the model was rarely the bottleneck anymore. The pipeline was.

FAQ

Was Deepgram Flux Multilingual the first code-switching STT model? It was the first single-model conversational STT with native code-switching. Older approaches used per-language models or language identification as a separate step. Flux handled language shifts within a single WebSocket stream with no external routing.

How did AssemblyAI’s Voice Agent API compare to Deepgram’s raw STT? They served different use cases. AssemblyAI’s Voice Agent API bundled STT, LLM, and TTS at $4.50/hr. Deepgram’s Flux STT was $0.39/hr streaming (at $0.0065/min) with no bundled orchestration. Teams that wanted a full voice pipeline out of the box picked AssemblyAI. Teams that wanted to compose their own stack picked Deepgram.

Could Google Cloud STT handle multilingual conversations? Google Chirp supported 100 plus languages but configured as language-per-channel rather than single-model code-switching. A developer building a multilingual voice agent on Google STT would need to handle language detection and channel switching in the application layer.

Why did Whisper not dominate the market in 2026? Whisper was competitive on accuracy and free to self-host, but had no official streaming endpoint, no turn detection, and no speaker diarisation in the open-weight version. The missing infrastructure layer made it expensive to use in production despite the zero licensing cost.

Was the STT accuracy gap actually measurable by June 2026? The gap between top-tier providers on clean English audio was measurable but not practically significant for most use cases. The difference showed up on accented speech, noisy environments, and domain-specific vocabulary, where training data composition mattered more than the model architecture.